Jolly-Seber models in MARK

Chapter 12 Jolly-Seber models in MARK Carl James Schwarz, Simon Fraser University A. Neil Arnason, University of Manitoba The original Jolly-Seber (JS) model (Jolly, 1965; Seber, 1965) was primarily interested in estimating abundance. Since then, the focus of many mark-recapture experiments changed to estimating survival rates (but not abundance) using the Cormack-Jolly-Seber (CJS) models (Cormack, 1964; Jolly, 1965; Seber, 1965) particularly with the publication of Lebreton et al. (1992). In previous chapters concerning analysis of live encounter data, we have focussed exclusively on CJS models. In recent years, however, interest has returned to estimating parameters related to abundance such as population growth (λ i ), recruitment ( f i ), as well as abundance (N i ). * Much of the theory about estimating population growth, recruitment, and abundance can be found in Williams et al. (2002). 12.1. Protocol The protocol for JS experiments is very similar to that of CJS experiments. In each of K sampling occasions, animals are captured. Unmarked animals are tagged with individually identifiable tags and released. Previous marked animals have their tag numbers read and are again released. The key difference between JS and CJS experiments is the process by which unmarked animals are captured and marked. In CJS experiment, no assumptions are made about how newly marked animals are obtained. The subsequent process of recovering marked animals in CJS models is conditional upon the animal being released alive at first encounter, and survival and catchability refer only to these marked animals. In JS experiments, the process by which unmarked animals are newly captured to be marked and released is crucial the assumptions about this process allows the experimenter to estimate recruitment and population sizes. In particular, it is assumed that unmarked animals in the population have the same probability of capture as marked animals in the population, i.e., that newly captured unmarked animals are a random sample of all unmarked animals in the population. * One of the reasons for preferring estimation of population growth is that estimates of population growth are fairly robust against heterogeneity in catchability (Schwarz, 2001), and tag loss (Rotella and Hines, 2005). Losses on capture are possible at every sampling occasion and are ignored in the discussion that follows. Of course, we hope that the survival of the marked subset of animals tells us something about the remaining unmarked animals in the population at large. Cooch & White (2013) 06.05.2013

12.2. Data 12-2 This assumption of equal catchability for marked and unmarked animals is needed to estimate abundance or recruitment or population growth and is required for the Pradel, Link-Barker, POPAN, and Burnham JS formulations in MARK... Other assumptions about the experiment are similar to those for the CJS model: Animals retain their tags throughout the experiment. * Tags are read properly. Sampling is instantaneous. Survival probabilities are the same for all animals (marked and unmarked) between each pair of sampling occasions (homogeneous survival). Catchablity is the same for all animals (marked and unmarked) at each sampling occasion (homogeneous catchability). This is the most crucial assumption for JS models. The study area is constant. If the study area changes over time, then the population size may change with the changing size of the study area. There are generally two sources of non-closure in any particular study. Animals may leave the population through death or permanently emigrate. Conversely, animals may enter the study area from outside (immigration) or be recruited from within the study area (e.g. fish growing into the catchable portion of the population). Specific tests for closure have been developed (e.g., Stanley and Burnham, 1999), but more often tests for closure are performed by fitting models with no apparent mortality (φ = 1), or no apparent recruitment ( f = 0,λ =φ, or b = 0), or both and letting the AIC c indicate the appropriate weight for such simpler models. 12.2. Data The basic unit of analysis is the capture history, a sequence of 0 s and 1 s that indicates when a particular animal was seen in the experiment. The JS models in MARK use the LLLLL capture history format. For example, the history (011010) indicates that an animal was captured for the first time at sampling occasion 2, was seen again at sampling occasion 3, not seen at sampling occasion 4, seen at sampling occasion 5, and not seen after sampling time 5. Either individual or grouped capture histories may be used. In many papers, the list of capture histories is too long to publish, and so a series of summary statistics are commonly used (Table 12.1; see also the reduced and full m-array descriptions in Chapter 5). For example, the history (011010) would contribute a count of 1 to n 2, n 3, n 5, u 2, m 3, m 5, R 2, R 3, R 5, r 2, r 3, and z 3. * Refer to Cowen and Schwarz (2006) for dealing with tag loss in JS experiments. Refer to Pledger and Efford (1998) for details on dealing with heterogeneity in JS models. Again losses on capture are ignored for now but are handled in the same way as elsewhere in MARK.

12.3. Multiple formulations of the same process 12-3 Table 12.1: Summary statistics often used for JS experiments. Losses-on-capture are found as n i R i. Statistics Definition n i Number of animals captured at occasion i, both marked and unmarked. n i = m i + u i. u i Number of unmarked animals captured at occasion i. m i Number of previously marked animals captured at occasion i. R i r i z i Number of animals released alive at occasion i +, i.e., just after sampling occasion i. Number of animals from R i that are subsequently captured after occasion i. Number of animals seen before i, seen after occasion i, but not seen at occasion i. While these summary statistics form the sufficient statistics for the Jolly-Seber probability model, their use has fallen out of favor in place of the raw histories used by MARK for two reasons. First, the use of individual covariates will require the individual capture history vectors (see Chapter 2, and Chapter 11). Second, it is difficult to compute goodness-of-fit statistics (i.e., the RELEASE suite of tests; Chapter 5) from the summary statistics. * If only summary statistics are available, it is possible to work backwards and create a set of histories that will reproduce these summary statistics that can be used with MARK to fit various models. One problem in using these pseudo-histories is that goodness-of-fit tests are nonsensical the goodness-of-fit tests require the full capture history of each animal. 12.3. Multiple formulations of the same process There are a number of formulations used in MARK to estimate abundance and related parameters, e.g., the POPAN; the Link-Barker and Pradel-recruitment; and the Burnham JS and Pradel-λ formulations. All of these models are slightly different parameterizations of the underlying population processes, and all are (asymptotically) equivalent in that they should give the same estimates of abundance and related parameters. The two main differences among the various formulations are 1. the way in which they parameterize new entrants to the population 2. if estimation is conditional upon the animals actually seen in the study (refer to Sanathanan 1972, 1977). All of the formulations model the recapture of marked animals in the same way. In this section, several of these models will be examined and contrasted. * Indeed, if the summary statistics are used by themselves, the fully time dependent JS models will be perfect fit to the summary statistics regardless if the model overall is a good fit.

12.3.1. The Original Jolly-Seber formulation 12-4 12.3.1. The Original Jolly-Seber formulation In the original JS formulation of Jolly (1965) and Seber (1965), the population process can be modeled as shown in Figure 12.1. The parameters p i andφ i are similar, but not identical to those in the CJS models. The parameter p i is the probability of capture of both unmarked and marked animals that are alive at occasion i (the CJS models referred only to marked animals); the parameterφ i refers to the survival probabilities of both marked and unmarked animals between occasions i and i + 1 (the CJS models referred only to marked animals). Figure 12.1: Original process model for JS experiments. p i represents the probability of capture at occasion i; φ i represents the probability of an animal surviving between occasions i and i + 1; and M i and U i represent the number of marked and unmarked animals alive at occasion i. Losses-on-capture are not modeled here, but are easily included. φ 1 φ 2 φ 3 φ 4 t 1 t 2 t 3 t 4 t 5... p 1 p 2 p 3 p 4 p 5 M 1 M 2 M 3 M 4 M 5 U 1 U 2 U 3 U 4 U 5 The number of marked animals in the population just before occasion i + 1 is found as M i+1 = (M i + u i )φ i where u i is the number of newly unmarked animals captured and subsequently marked. The number of net new entrants to the population was defined as B i = U i+1 φ i (U i u i ) The B i values refer to the net number of new entrants to the population between sampling occasion i and occasion i +1. The reference to net number of new entrants implies that animals that enter between two sampling occasions but then die before being subject to capture at occasion i+1 are excluded. * As in the CJS models, the term survival refers to apparent survival permanent emigration is indistinguishable and treated the same as mortality. Similarly, the term births refers to any new animals that enter the study population regardless if in situ natural births or immigration from outside the study area. The likelihood function consists of three parts. The first part models losses-on-capture using a simple binomial distribution as in the CJS models. The second part models the recapture of marked animals in exactly the same way as in the CJS model. Finally, the third part models the number of unmarked animals captured at occasion i as a binomial function of the number of unmarked animals in the population, i.e., u i is Binomial(U i, p i ). The estimates of p i andφ i are found in exactly the same way as in the CJS models. The estimated unmarked population sizes were estimated as Û i = u i / ˆp i. The estimated number of births was found by substituting in the estimates in the previous definition and does not form part of the likelihood. Finally, the estimates of population size at each time point is found by adding the estimates of U i and M i. * The term gross number of entrants would include these deaths prior to the next sampling occasion. Refer to Schwarz et al. (1993) for details on the estimation of these gross births.

12.3.2. POPAN formulation 12-5 12.3.2. POPAN formulation Schwarz and Arnason (1996) adopted a slightly different parameterization for a number of reasons: The parameters B i never directly entered into the likelihood function. The number of entrants must be non-negative, but it was difficult to enforce ˆB i 0 and negative estimates of births were often obtained. Because the B i did not appear in the likelihood, how could these be forced to be equal across groups following the Lebreton et al. (1992) framework? How could death-only models (e.g., all B i known to be zero) or birth-only models (allφ i =1) or closed models be obtained by constraining the likelihood function. In their parameterization, first implemented in the computer package POPAN and now a sub-module of MARK, they postulated the existence of a super-population consisting of all animals that would ever be born to the population, and parameters b i which represented the probability that an animal from this hypothetical super-population would enter the population between occasion i and i + 1 as shown in Figure 12.2. * Figure 12.2: Process model for POPAN parameterization of JS experiments. p i represents the probability of capture at occasion i;φ i represents the probability of an animal surviving between occasions i and i + 1; and b i represents the probability that an animal from the super-population (N) would enter the population between occasions i and i + 1 and survive to the next sampling occasion i + 1. Losses-on-capture are assumed not to happened, but are easily included. b 0 b 1 b 2 b 3 b 4 N φ 1 φ 2 φ 3 φ 4 t 1 t 2 t 3 t 4 t 5... p 1 p 2 p 3 p 4 p 5 Now the expected number of net new entrants is simply found as E[B i ] = Nb i. If B 0 represents the number of animals alive just prior to the first sampling occasion, then N = B 0 + B 1 + B 2 + + B K 1 i.e., the total number of animals that ever are present in the study population. The parameters b i are referred to as PENT (Probability of Entrance) probabilities in MARK. Notice that b 0 +b 1 + +b K 1 = 1; this will have consequences later when the models are fitted using MARK. Even though the number of new animals is not modeled in the process, modeling the entrance probabilities and a super-population size is equivalent. * The super-population approach was first described by Crosbie and Manly (1985) where distribution functions (e.g. a Weibull distribution) was used to model survival time once an animal had entered the population. To our knowledge, there is no readily available computer code for the Crosbie and Manly (1985) model.

12.3.2. POPAN formulation 12-6 Under this parametrization, E[N 1 ] = Nb 0 E[N 2 ] = E[N 1 ]φ 1 + Nb 1. The probability of any capture history can be expressed using these parameters. For example, P[(01010)] is found as: P [(01010)] = [ b 0 ( 1 p1 ) φ1 + b 1 ] p2 φ 2 ( 1 p3 ) φ3 p 4 [ 1 φ4 +φ 4 ( 1 p5 )] As in the CJS models, the fate of the animal after the last capture is unknown either it died, or it survived and was not seen at occasion 5. In a symmetrical fashion, the fate of the animal before the first time it is captured is also unknown. Either it was present in the population prior to sampling occasion 1 and wasn t seen at occasion 1 and survived to occasion 2, or it entered the study population between sampling occasions 1 and 2 and survived to sampling occasion 2 where it was captured for the first time. The likelihood function is again a multinomial function over all the observed capture histories. Schwarz and Arnason (1996) showed that it could be factored into three parts: L = P ( first capture ) P ( subsequent recaptures ) P ( loss on capture ) where the second and third components are identical to the CJS models. It turns out that similar to CJS models, not all parameters are identifiable and only functions of parameters can be estimated in the fully time-dependent model. The set of non-identifiable parameters is given in Table 12.2. In particular, the final survival and catchability parameters are confounded (as in the CJS models), and symmetrically the initial entrance and catchability are confounded. This impacts three other sets of parameters, in particular N 1 and N K cannot be cleanly estimated, nor can b 1 and b K 1. If confounding takes place, the estimated super-population number may be suspect, so some care must be taken in fitting appropriate models. For example, models with equal catchability over sampling occasions make all parameters identifiable. This confounding implies that careful parameter counting may have to be done when fitting POPAN models. The fully time-dependent model{p t,φ t, b t } has K parameters for catchability, K 1 parameters for survival, K parameters for the PENTs, and 1 parameter for the super-population size for a total of 3K parameters. However, not all are identifiable and the PENTs must sum to one. Only the products b 0 p 1 andφ K 1 p K can be estimated, and one of the PENTs is not free (as the sum must equal 1), leaving 3K 3 parameters that can be estimated for each group. Furthermore, as indicated in Table 12.2, the b 1 and b K 1 parameters are affected (the estimates reflect the combination of parameters as listed in the table) which further affect N 1 and N K. While the latter parameter combinations are estimable, they seldom represent anything biologically useful. The actual number of parameters reported by MARK in the results browser should be checked carefully.

12.3.3. Link-Barker and Pradel-recruitment formulations 12-7 Table 12.2: Confounded parameters in the POPAN parameterization in the fully time-dependent model. In order to resolve this confounding, the models must make assumptions about the initial (p 1 ) and final (p K ) catchabilities. For example, a model may assume that catchabilities are equal across all sampling occasions. Function φ K 1 p K b 0 p 1 b 1 + b 0 (1 p 1 )φ 1 b K 1 /φ K 1 Interpretation Final survival and catchability. Initial entrance and catchability. Entry between first and second occasions cannot be cleanly estimated because initial entrance probability cannot be estimated. MARK (and other programs) will report an estimate for this complicated function of parameters but it may not be biologically meaningful. Entry prior to last sampling occasion cannot be cleanly estimated because final survival probability cannot be estimated. MARK (and other programs) will report an estimate for this complicated function of parameters but it may not be biologically meaningful. Once estimates of p,φ, b, and N are obtained, the estimated number of births is obtained as ˆB i = ˆN ˆb i. The estimated population sizes are obtained in an iterative fashion: ˆN 1 = ˆB 0 ˆN 2 = ˆN 1 φ 1 + ˆB 1. If losses on capture occur, they are removed before the population size at occasion i is propagated to occasion i + 1. The likelihood does not contain any terms for B i or N i these are derived parameters and standard errors for these estimates are found using the Delta method (see Appendix 2). 12.3.3. Link-Barker and Pradel-recruitment formulations The Link-Barker (2005) and Pradel-recruitment * (1996) formulations are conceptually the same and the process model is shown in Figure 12.3. The parameters for survival (φ i ) and catchability (p i ) are standard. The parameter f i is interpreted as a per capita recruitment probability, i.e., how many net new animals per animal alive at occasion i enter the population between occasion i and i + 1? Unlike the POPAN formulation, the Link-Barker formulation conditions upon an animal being seen somewhere in the experiment. This eliminates the necessity of estimating the super-population size, but also means that abundance cannot be directly estimated. Any probability of a history must be normalized by the probability of being a non-zero history. * There are three different Pradel models and -recruitment refers to the Pradel models parameterized using f i terms

12.3.3. Link-Barker and Pradel-recruitment formulations 12-8 Figure 12.3: Process model for Link-Barker and Pradel-recruitment parameterization of JS experiments. p i represents the probability of capture at occasion i;φ i represents the probability of an animal surviving between occasions i and i + 1; and f i represents the net recruitment probability, i.e., the per capita number of new animals that enter between occasions i and i + 1 and survive to the next sampling occasion i + 1 per animal alive at occasion i. Losses-on-capture are assumed not to have happened, but are easily included. f 1 f 2 f 3 f 4 φ 1 φ 2 φ 3 φ 4 t 1 t 2 t 3 t 4 t 5... p 1 p 2 p 3 p 4 p 5 For example, P[(01010) animal seen] is proportional to: P [(01010) animal seen] [ (1 p 1 )φ 1 + f 1 ] /p1 p 2 φ 2 ( 1 p3 ) φ3 p 4 [ 1 φ4 +φ 4 ( 1 p5 )] where the constant of proportionality is related to the probability of seeing an animal somewhere in the experiment. As in the CJS models, the fate of the animal after the last capture is unknown either it died, or it survived and was not seen at occasion 5. In a symmetrical fashion, the fate of the animal before the first time it is captured is also unknown either it was present among animals seen in the experiment at time 1, was not seen, and survived to time 2, or it entered between times 1 and 2. The likelihood function is a multinomial function over all the observed capture histories conditional upon an animal being seen somewhere in the experiment. The implementation of the Link-Barker model differs from the Pradel-recruitment formulation in a number of ways. First, Link-Barker partitioned the likelihood in a similar fashion to the POPAN formulation which made it easier to implement in the Bayesian context of their paper. Second, Link- Barker explicitly modeled the confounded parameters (but the MARK implementation leaves the confounded parameters separate and it is the user s responsibility to understand the confounding). Third, losses on capture are handled differently between the two formulations and this affects the interpretation of the recruitment parameters. The Link-Barker model also differs from the POPAN formulation as there is no need to postulate the existence of a super-population the model is fit conditional upon the observed number of animals in the experiment. * Section 12.3.5 outlines the equivalences between the Link-Barker parameters and those of other formulations. As in the POPAN formulation, the fully time-dependent Link-Barker and Pradel-recruitment models have a number of parameter confoundings as listed in Table 12.3. On the surface, the fully timedependent model {p t,φ t, f t } has K catchability parameters, K 1 survival parameters, and K 1 recruitment parameters for a total of 3K 2 parameters. However, only the productφ K 1 p K and the ratio f 1 /p 1 can be estimated, leaving a net of 3K 4 parameters. The estimate for f K 1 estimates a function of other parameters as shown in Table 12.3. * The implementation of Schwarz and Arnason (1996) in the POPAN package also estimates parameters conditional upon being seen, and then adds another step to estimate the super-population size.

12.3.4. Burnham JS and Pradel-λ formulations 12-9 Table 12.3: Confounded parameters in the Link-Barker parameterization in the fully time-dependent model. In order to resolve this confounding, the models must make assumptions about the initial (p 1 ) and final (p K ) catchabilities. For example, a model may assume that catchabilities are equal across all sampling occasions. Function Interpretation φ K 1 p K Final survival and catchability (φ 1 + f 1 )/p 1 Initial recruitment and survival f K 1 p K Final recruitment and catchability cannot be cleanly estimated. MARK (and other programs) will report an estimate for this complicated function of parameters but it may not be biologically meaningful. The abundance at each sampling occasion and the absolute number of new entrants cannot be estimated even as derived parameters because of the conditioning upon animals seen at least once during the experiment. 12.3.4. Burnham JS and Pradel-λ formulations The final formulations to be considered in this chapter model new entrants to the population indirectly by modeling the rate of population growth (λ) between each interval where population growth is the net effect of survival and recruitment. Ifφ i is the decrease in the population per member alive at time i, and f i is the increase in the population per member alive at time i, then the sum of their contributions is the net population growth: λ i = N i+1 /N i =φ i + f i These formulations were developed by Burnham (1991) and Pradel (1996). The key difference between the two parameterizations is that the Pradel-λ approach is conditional upon animals being seen during the study, while the Burnham JS formulation is not. Therefore, the Burnham Jolly-Seber formulation also includes a parameter for the population size at the start of the experiment. This enables the estimation of the population size at each subsequent time point. However, in practice, it is often difficult to get the Burnham-JS model to convergence during the numerical maximization of the likelihood. Although the implementation of this model has been thoroughly checked and found to be correct, MARK has some difficulty obtaining numerical solutions for the parameters because of the penalty constraints required to keep the parameters consistent with each other. * For this reason, only the Pradel-λ formulation will be discussed further in this section (and treated in depth in Chapter 13). The process model is shown in Figure 12.4. The parameters for survival (φ i ) and catchability (p i ) are standard. The parameterλ i is interpreted as the ratio of successive population abundances. * This convergence problem disappears for some models if simulated annealing is used for the numerical optimization. - J. Laake & E. G. Cooch, pers. obs.

12.3.4. Burnham JS and Pradel-λ formulations 12-10 Figure 12.4: Process model for Burnham and Pradel-λ parameterization of JS experiments. p i represents the probability of capture at occasion i;φ i represents the probability of an animal surviving between occasions i and i + 1; andλ i represents the rate of population change. The population size at time 1, N 1 is used by the Burnham formulation, but not by the Pradel-λ formulation. Losses-on-capture are assumed not to happened, but are easily included. φ 1 φ 2 φ 3 φ 4 λ 1 λ 2 λ 3 λ 4 t 1 t 2 t 3 t 4 t 5... p 1 p 2 p 3 p 4 p 5 N 1 Unlike the POPAN and Burnham formulations, the Pradel-λ formulation conditions upon an animal being seen somewhere in the experiment. This eliminates the necessity of estimating the population sizes at any sampling occasion. But, the probability of a history must now be normalized by the probability of being a non-zero history. For example, P[(01010) animal seen] is proportional to: P [(01010) animal seen] [ λ 1 p 1 φ 1 ] p2 φ 2 ( 1 p3 ) φ3 p 4 [ 1 φ4 +φ 4 ( 1 p5 )] where the constant of proportionality is related to the probability of seeing an animal somewhere in the experiment. As in the CJS models, the fate of the animal after the last capture is unknown either it died, or it survived and was not seen at occasion 5. In a symmetrical fashion, the fate of the animal before the first time it is captured is also unknown either it was present at the initial sampling occasion and not seen, or was part of the population growth (over and above survival from the first sampling occasion). The likelihood function is a multinomial function over all the observed capture histories conditional upon an animal being seen somewhere in the experiment. Section 12.3.5 outlines the equivalences between the Pradel-λ parameters and those of other formulations. As in the POPAN formulation, the fully time-dependent Pradel-λ formulation has a number of parameter confoundings as listed in Table 12.4. On the surface,the fully time-dependent model{p t,φ t,λ t } has K catchability parameters, K 1 survival parameters, and K 1 growth parameters for a total of 3K 2 parameters. However, only the productφ K 1 p K and the functionλ 1 φ 1 p 1 can be estimated, leaving a net of 3K 4parameters. Furthermore, the estimate forλ K 1 estimates a function of other parameters and may not be interpretable. The abundance at each sampling occasion and the absolute number of new entrants cannot be estimated even as derived parameters because of the conditioning upon animals seen at least once during the experiment.

12.3.5. Choosing among the formulations 12-11 Table 12.4: Confounded parameters in the Pradel-λ parameterization in the fully time-dependent model. In order to resolve this confounding, the models must make assumptions about the initial (p 1 ) and final (p K ) catchabilities. For example, a model may assume that catchabilities are equal across all sampling occasions. Function φ K 1 p K λ 1 φ 1 p 1 λ K 1 p K Interpretation Final survival and catchability Initial growth and survival Final recruitment and catchability cannot be cleanly estimated. MARK (and other programs) will report an estimate for this complicated function of parameters but it may not be biologically meaningful. Because population growth (λ) is a function both of survival and recruitment (i.e.,λ i = φ i + f i ), the modeler should be careful about fitting simpler models that restrict population growth but leave survival time-dependent. For example the model { p t,φ t,λ } would imply that recruitment varies in a time dependent fashion to exactly balance changes in survival to keep population growth constant. This may not be a sensible biological model. Another potential problem with the Pradel-λ model is that no constraints are imposed in MARK that population growth must exceed the estimated survival probability. Consequently (as seen in the examples that follow), it is possible to get estimated survival probabilities of 80% while the estimated population growth rate is only 70%. This logically cannot happen, and is often an indication that recruitment did not occur in that interval the illogical estimates are artifacts of the estimation process. Under these circumstances, models that separate recruitment from population growth may be preferred. 12.3.5. Choosing among the formulations All of the formulations use the same input file in the same format. Which JS formulation should be used for a particular experiment? There are two considerations. First, only certain of the formulations can be used in MARK if losses-on-capture occur in the experiment. Secondly, and more importantly, different formulations give you different types of information and can be used to test different hypotheses. All of the formulations should give the same estimates of survival and catchability as all formulations estimate these from recaptures of previously marked animals using a CJS likelihood component. Even though all the models give different types of estimates for growth or recruitment or births, it is always possible to transform the estimates from one type to another by simple transformation and the standard errors can be found using the delta method. The major equivalents are between NET births, recruitment, and population growth parameters. Recruitment parameters are the net number of new animals that enter the population between occasions i and i + 1 per animal present in the population at occasion i f i = B i /N i = Nb i /N i

12.3.5. Choosing among the formulations 12-12 Population growth is the proportionate increase in abundance between occasions i and i + 1: λ i = N i+1 N i = N i φ i + B i N i =φ i + f i Actual estimates from fitted models may not follow these exact relationships for several reasons. First, certain estimates (e.g., apparent survival) should be constrained to lie between 0 and 1. If an estimated survival hits against this boundary, estimates of survival prior to and after this sampling occasion will also be affected. If the identity link of MARK is used, estimates are allowed to fall outside these normal boundaries, and usually the exact relationships above then hold true among the estimates as well. Second, losses on capture complicate the equivalences among parameters. For example, Link and Barker (2005) indicate that their f i should be interpreted as the number of new animals that enter between occasions i and i + 1 per hypothetical animals alive at occasion i in the absence of losses on capture, while the Pradel-recruitment f i is the number of new animals after losses on capture have been taken into account. Lastly, MARK does not impose constraints that estimated population growth parameters must be at least as great as estimated survival probabilities. Consequently, it is possible (as seen in the examples) that the estimated population growth rate is less than the estimated survival probability which would imply a negative recruitment. In my opinion, formulations that model recruitment and survival as separate processes are preferred in these case the estimated population growth rate can always be derived from these alternative models. It cannot be emphasized too strongly, that all of the formulations require the same careful attention to study design in particular the study area must remain a consistent size, and the probability of capturing an unmarked individual must be the same as a marked individual at each sampling occasion. A Cormack-Jolly-Seber experiment where marked animals are captured and released haphazardly, should not be then analyzed using any of the formulations of the Jolly-Seber model discussed in this chapter. Table 12.5: Summary of criteria to choose among the different JS formulations losses on estimates available for formulation capture abundance net births recruitment λ POPAN yes yes yes no no Link-Barker yes no no yes yes Pradel-recruitment no no no yes yes Burnham JS yes yes yes no yes Pradel-λ yes no no no yes The implementation of Burnham s JS model in MARK often does not converge, and is not recommended (although convergence problems may be minimized for some models if simulated annealing is used for the numerical optimization. - J. Laake & E.G. Cooch, pers. obs.) The standalone package of POPAN will estimate recruitment, and population growth as derived parameters

12.3.6. Interesting tidbits 12-13 12.3.6. Interesting tidbits There have been a number of queries in the MARK forum (http://www.phidot.org/forum) about the use of POPAN and other models to estimate abundance. This section will try to answer some of these queries in more detail. Deviance of 0 in POPAN The following query was received in the MARK forum (http://www.phidot.org/forum, 2006-04-07) which asked: I read on one of the other posts that getting a deviance of zero was possible using the robust design model because the saturated model hadn t been computed yet and some constants were left out for faster computation. Is something along these lines at play in POPAN also because when I run a particular set of data, I get a deviance of zero? The deviance of models is often computed as the negative of twice the difference in the log-likelihood between the current model and a saturated model. The usual saturated model in CJS and other models that condition upon an animal s first capture, is to have a separate probability for each observed history, i.e., ifωis a capture history (e.g.,010011) then the P(ω) under the saturated model is found as P(ω) = n ω n obs where n ω is the number of animals with capture historyω, and n obs is the total number of animals observed. In such cases, the log-likelihood of the saturated model (ignoring constants) is then n ω log ( ) ( nω ) p ω = n ω log n obs However, this doesn t work for models where abundance is estimated because you need to include the animals with history (000000...), i.e., those animals not observed. This can only be computed if the population size is estimated which cannot be estimated under the saturated model. Hence a deviance cannot be directly computed. However, the likelihood can be portioned as shown earlier into components representing the probability of first capture, the probability of subsequent recapture given the animal has been captured, and the probability of losses-on-capture given that an animal is captured. Schwarz and Arnason (1996) and Link and Barker (2005) showed that the first component is essentially non-informative about the capture, survival, and loss-on-capture rates. This suggests that an approximate deviance could be computed using only the latter two components, i.e., by conditioning upon animals that are seen at least once in the experiment. The difference between the two likelihoods would be based only on the part representing animals seen at least once. In practice, we would suggest that you compute a deviance based on conditioning on the observed animals. The easiest way is to use the Link-Barker model (which doesn t estimate abundance) but is equivalent to the POPAN model. So if you fit a {p t,φ t, b t } model in POPAN look at the deviance of the {p t,φ t, f t } model in Link-Barker formulation. This could be used to estimate a variance-inflation factor to adjust reported standard errors. 12.4. Example 1 - estimating the number of spawning salmon After spending several years at sea, coho salmon (Oncorhynchus kisutch) return to spawn in the Chase River, British Columbia. The normal life cycle of coho salmon is to return at age 3 as adults to spawn

12.4. Example 1 - estimating the number of spawning salmon 12-14 and die. But, some precocious males return earlier at age 2 to spawn and die. One question of interest is if the distribution of salmon that return to spawn at different parts in the spawning period the same for regular adult and precocious males? As fish return to the Chase River, they are captured using electrofishing gear. If they are unmarked they are given a unique tag number and released. If they were previously marked, the tag number is read. The experiment took place over a 10 week period in 1989, but the data from weeks 1 and 2, and weeks 9 and 10 were pooled and labeled as weeks 1.5 and 9.5. Approximately the same amount of effort was expended in each week of sampling. More details of this experiment are found in Schwarz et al. (1993). The datafile is given in thechase.inp file in the sample data directory and a portion of it is reproduced in Figure 12.5. This study has two groups, the regular adult males and the precocious males (called jacks). Notice that a separate capture history is used for each tagged fish - unfortunately, this implies that the residual plots and deviance plots in MARK cannot be used to assess goodness of fit. Figure 12.5: Portion of the data for the Chase 1989 experiment /* Estimating salmon numbers returning to spawn in Chase River 1989 */ /* These are the male salmon with two groups. */ /* Group1 = adults. group2=jacks */ /* Survey conducted over 10 weeks. Weeks 1 & 2 pooled. weeks 9 & 10 pooled */ 11000000 1 0 ; /* tagnum=1 */ 10000000 0 1 ; /* tagnum=3 */ 10000000 0 1 ; /* tagnum=4 */ 10000000 0 1 ; /* tagnum=6 */ 11000000 0 1 ; /* tagnum=8 */ 10110000 0 1 ; /* tagnum=9 */ 10000000 0 1 ; /* tagnum=10 */ 10000000 1 0 ; /* tagnum=11 */... additional histories follow... Summary statistics for this experiment are presented in Table 12.6. Table 12.6: Summary statistics for the Chase 1989 experiment statistics for adults statistics for jacks t i n i m i u i R i r i z i t i n i m i u i R i r i z i 1.5 37 0 37 37 12 0 1.5 67 0 67 62 21 0 3.0 22 6 16 21 13 6 3.0 28 9 19 25 7 12 4.0 52 7 45 41 19 12 4.0 46 6 40 44 9 13 5.0 56 17 39 54 26 14 5.0 47 12 35 45 5 10 6.0 46 26 20 38 8 14 6.0 25 9 16 24 3 6 7.0 28 16 12 20 2 6 7.0 16 6 10 12 1 3 8.0 22 3 19 16 2 5 8.0 7 1 6 5 1 3 9.5 10 7 3 0 0 0 9.5 7 4 3 0 0 0

12.4.1. POPAN formulation 12-15 Because n i > R i for some sampling occasions, this indicates that some losses-on-capture occurred. For example, there was one adult loss-on-capture in week 3; 11 adults lost in week 4, etc. 12.4.1. POPAN formulation The POPAN super-population is a natural way to think of this experiment - a pool of fish is returning to spawn. During each week, a certain fraction of these returning fish decide to enter the spawning areas. Let us begin by fitting a model only to the regular adults, i.e., to the first group. Later models will be fit to both groups. Launch MARK. Select the POPAN module, enter the number of sampling occasions and the number of attribute groups: Don t forget to set the intervals between sampling occasions. As weeks (1 and 2) were pooled and labeled as week 1.5, the interval between the first and second sampling occasion is 1.5 weeks. Similarly weeks (9 and 10) were pooled and so the last interval is also longer.

12.4.1. POPAN formulation 12-16 Start by fitting a fully-time dependent model {p t,φ t, b t } (using an obvious notation extension from CJS models). In this model there are 8 sampling occasions which give rise to 8 capture probabilities, 7 apparent survival probabilities, 8 probability of entry probabilities, and 1 super-population parameter. Note that MARK does not allow the user to specify the parameter b 0 (the proportion of the population available just before the first sampling occasion) and so only presents 7 PENT parameters in the PIM and output corresponding to b 1,...,b 7. * Note that unlike the CJS models, we assume a single set of survival and catchability parameters, regardless of when previously captured. The super-population size has its own parameter. Because b 0 + b 1 + + b K 1 = 1, a special link-function must be specified for the PENT parameters. This is done using the Parameter specific link function radio button: * In the stand alone POPAN package, the user has access to all of the PENT s.

12.4.1. POPAN formulation 12-17 The sin or logit or any of the other link functions can be used for the p andφparameters. In order to specify that a set of parameters must sum to 1, the Multinomial Logit link function (called Mlogit in MARK) must be used. If there are several groups, each set of PENTs must independently sum to 1, so MARK provides several sets of MLogit link functions. As there is only one group, the MLogit(1) linkfunction is used for the PENTs. It is possible to specify that some of the PENTs are zero if, for example, the experimenter knew that no new animals entered the study population during this interval. Also notice,that a log or identity link should be used for the super-population size as it is not restricted to lie between 0 and 1: Now run MARK. If we look at the REAL estimates, we must keep in mind that not all parameters are identifiable:

12.4.1. POPAN formulation 12-18 In particular, the final survival and catchability are confounded as in the CJS model. The initial entrance and catchability parameters are also confounded (only the product b 0 p 1 can be estimated) however, MARK does not report b 0 so the first PENT reported here refers to b 1 which cannot be estimated separately (refer to Table 12.2). This non-identifiability can often be recognized by the large standard errors for certain estimates, or by estimates tending to the value 1.0. Also notice, that because the intervals are unequal size, the survival probabilities are given on a per week basis, so that the survival probability for the initial 1.5 week interval is found as 0.57 1.5 = 0.43. The estimates of population size and net births are found under the derived parameter section: Again, not all parameters are identifiable. The derived parameters also include estimates of gross births - these are explained in more detail in Schwarz et al. (1993). Goodness-of-fit can be assessed using the RELEASE suite as in CJS models (see Chapter 5 for details on RELEASE). The results are shown below.

12.4.1. POPAN formulation 12-19 There is some evidence of potential lack-of-fit as indicated by Test2 in component C3, but a detailed investigation of that table shows it is not serious. Unfortunately, because individual capture histories were used, residual plots are not useful. Because of parameter confounding, it is important to count the actual number of estimable parameters. On the surface there are 24 parameters composed of 8 capture parameters, 7 survival parameters, 8 PENT parameters and 1 super-population parameter. However, the PENTs must sum to 1, and Table 12.2 indicates that two parameters are lost to confounding which leaves a net of 21 actual identifiable parameters. Check that the results browser shows 21 parameters for this model. The original sampling experiment has approximately equal effort at all sampling occasions. Perhaps a model with constant catchability over time is suitable, i.e., model {p,φ t, b t }. This model is specified in the PIM in the usual fashion: None of the other PIM s need to be respecified. The model is run, and again the parameter specific link functions must be specified for the PENTs (use the MLogit(1) link function) and for the super-population size (use the log link function): Now all parameters are identifiable (shown in the following two figures at the top of the next page). The results show ˆb 1 = 0.044, ˆb 2 = 0.33, etc. These are interpreted as 4.4% of adult returning salmon return between weeks 1.5 and 3; about 33% of adult returning salmon return to spawn between weeks 4 and 5, etc. The value of ˆb 0 = 0.352 is obtained by subtraction (ˆb 0 = 1 ˆb 1 ˆb 2 ˆb K 1 ). This is interpreted as 35% of adults returning salmon has returned to spawn before sampling began in week 1.5. The total number of salmon returning to spawn (the super-population) is estimated to be ˆN = 332 (SE=29) fish. The derived birth parameters are found as ˆB i = ˆN ˆb i. For example, ˆB 1 = ˆN ˆb 1 = (332 0.044) = 14.9. This is interpreted as about 15 adult fish returned to spawn between weeks 1.5 and 3. ˆB 0 = ˆN ˆb 0 = (332 0.352) = 117 fish are estimated to be present before the first sampling occasion.

12.4.1. POPAN formulation 12-20 The derived estimates of population size are found iteratively and must account for losses-on-capture and the unequal time intervals (which affects the survival terms). ˆN 1 = ˆB 0 = 117.03. ˆφ 1.5 ˆN 2 = ( ˆN 1 loss 1 ) 1 + ˆB 1 = (117.03 0)0.584 1.5 + 14.92 = 67.2. 1.0 ˆN 3 = ( ˆN 2 loss 2 ) ˆφ 2 + ˆB 2 = (67.2 1)0.926 1.0 + 110.39 = 171.72. The number of identifiable parameters for this model is 16 composed of 1 capture parameters, 7 survival parameters, 8 PENT parameters and 1 super-population parameter less the restriction that the PENTs must sum to 1. The number of parameters reported in the browser may have to be manually adjusted to indicate the correct number of parameters. Another sub-model can also be fit where the apparent survival probability (per unit time) is constant over all intervals, i.e., model {p,φ, b t }. It is fit in the same fashion by adjusting the PIMs: This model would have 10 parameters composed of 1 capture parameter, 1 survival parameter, 8 PENT parameters, and 1 super-population parameters with 1 restriction that the PENTs sum to 1. The final results table (after making sure that the number of parameters is correct):

12.4.1. POPAN formulation 12-21 shows not much support for this final model. If model averaging is to be used, some care must be taken as not all parameters are identifiable in all models. For example, most models with a p t structure, will be unable to estimate abundance at the first (N 1 ) or last (N K ) sampling occasion; nor can recruitment be estimated for the first (b 1, or B 1 ) or last (b K 1 or B K 1 ) interval. Other models could be fit, e.g., an equal fraction of the super-population returns to spawn in each week, but in this example is highly unlikely biologically and was not fit. Now let us return to the real question of interest - do jacks and adults have the same return pattern? This is now a two group problem and is handled in a similar fashion to the ordinary CJS model. Start a new project and this time specify two groups (regular adults and jacks) rather than a single group. Also specify the names of the two groups. Now each parameter can vary over sampling occasions (intervals) and/or groups. The full model fit will be specified by a triplet of specifications. In the previous example, a model with equal catchability over sampling occasions was tenable, but adults and jacks may have different catchabilities. Survival probabilities varied by time, so perhaps start with a fully time- and group-dependent model forφ. Also start with a full group and time dependence for the PENTs. This would correspond to model {p g,φ g t, b g t } with the following PIMs.

12.4.1. POPAN formulation 12-22 The two PIMs for the super-population size of the adults and jacks (respectively) are not shown. Request that MARK fit this model. As before, we must indicate to MARK that the PENTs sum to 1, for each group using the parameter specific link functions:. There are a total of 8 sampling occasions, 8 PENTs per group, but MARK only shows the last seven (b 0 for each group is implicitly assumed). Use the MLogit(1) link function for the first seven PENTs (belonging to group 1) and the MLogit(2) link function for the last seven PENTs (belonging to group 2). Notice that only 30 parameters are displayed per window, so the More button must be used to scroll to the next page: Don t forget to specify that the two super-population parameters should have the identity or log link function.

12.4.1. POPAN formulation 12-23 Run the model and add it to the results browser. Then, run and fit the models {p g,φ t, b g t } and and {p g,φ t, b t }. * Count parameters carefully. All of the models have a simple structure for the p s so there is no problem with confounding. The model {p g,φ t, b g t } has 2 catchability parameters, 7 survival parameters, 16 PENTs, and 2 super-population sizes. However, each set of PENTs must sum to 1, leaving 25 free parameters. The model {p g,φ t, b t } has 2 catchability parameters, 7 survival parameters, 8 PENTs (but these must sum to 1), and 2 super-population parameters for a total of 18 parameters. The model {p g,φ g t, b g t } has 2 catchability parameters, 14 survival parameters, 18 PENTs (but each of the two groups of PENTs must sum to 1), and 2 super-population sizes for a total of 32 parameters. The final results window looks like (after adjusting for the number of parameters) The final models of interest are a comparison between model {p g,φ t, b g t } and {p g,φ t, b t } (why?). The AIC c shows very little support for the model where the jacks enter in the same distribution as the adults. In particular, examine the estimates of the PENTs for the {p g,φ t, b g t } model: Recall that even though there are 8 PENT parameters per group that MARK does not let you specify the b 0 parameter in the PIMS this value is obtained by subtraction from 1. Further manipulations can be done by copying the real parameter values to an Excel spreadsheet: * The Initial Copy 1 PIM to another PIM is helpful here.

12.4.1. POPAN formulation 12-24 You will find that about 34% (1 0.033 0.326... 0.000) of adults were in the stream prior to the first sampling occasion, while about 69% (1 0.000 0.227 0.036... 0.000) of jacks were in the stream prior to the first sampling occasion. So it appears that precocious males tend to return earlier (for this stream and year) than regular adults. In this example, it is vitally important that catchability be approximately constant across all sampling occasions so that a model with p g could be fit; any model where catchability varied across time (a p t or p g t model) would have the first PENT parameter hopelessly confounded with the first catchability parameter and in many cases, makes it difficult if not impossible to do sensible model comparisons. This again illustrates the need for careful study design with JS models. The multinomial logit link and the POPAN model begin sidebar In situations where you want to constrain estimates from a set of 2 or more parameters to sum to 1, you might use the multinomial logit link (MLogit), which was introduced in some detail in Chapter 10 with respect to multi-state models (where the transitions from a given stratum must logically sum to 1.0). While specifying the MLogit link in MARK is straightforward, you need to be somewhat careful. Consider the following set of parameters in a PIM for the probability of entry (pent) in a POPAN model: 61 62 63 64 65 66 67 68 69 70 The parameter-specific link would be selected in the Setup Numerical Estimation Run window, and the MLogit(1) link would be applied to parameters 61 70 to force these 10 estimates to sum to 1. But suppose that you wanted to force all of the 10 entry probabilities to be the same, and have the sum of all 10 be 1? You might be tempted to specify a PIM such as 61 61 61 61 61 61 61 61 61 61 (i.e., simply use the same index value for all the parameters in the PIM), but that would be incorrect. Changing the PIM and selecting the MLogit link for parameter 61 would result in parameter 61 alone summing to 1 (i.e., just like a logit link), but would not force the sum of the 10 values of parameter 61 to sum to 1. To implement the proposed model, the PIM should not be changed from the top example (i.e., it should maintain the indexing from 61 70), and the design matrix should be used to force the same estimate for parameters 61 70: Parameter Design Matrix 61 1 62 1 63 1 64 1 65 1 66 1 67 1 68 1 69 1 70 1 Then the MLogit(1) link should be specified for the 10 parameters 61 70. The result is that now all 10 parameters have the same value, and 10 times this value is le1. Another example - suppose you wanted parameters 61 and 62 to be the same value, 63 to 66 the same, 67 and 68 the same, and 69 and 70 the same, but the sum over all parameters to be 1. Again you would use the PIM

12.4.2. Link-Barker and Pradel-recruitment formulations 12-25 61 62 63 64 65 66 67 68 69 70 but again use the design matrix to implement the constraints. The following design matrix is one example that would produce such a set of constraints. Parameter Design Matrix 61 1 1 0 0 62 1 1 0 0 63 1 0 1 0 64 1 0 1 0 65 1 0 1 0 66 1 0 1 0 67 1 0 0 1 68 1 0 0 1 69 1 0 0 0 70 1 0 0 0 The key point with these examples is that the PIM cannot be used to constrain parameters if you want the entire set of parameters to sum to 1. Rather, the design matrix has to be used to make the constraints, with each of the entries in the PIM given the same MLogit(x) link. Further examples of the MLogit link are discussed in Chapter 10. end sidebar 12.4.2. Link-Barker and Pradel-recruitment formulations The Link-Barker or Pradel-recruitment formulation can be conveniently obtained by switching data types from any of the JS formulations. We will illustrate the use of the Link-Barker formulation; that for the Pradel-recruitment is similar, but in this case cannot be used because of losses-on-capture. Let us begin with a time dependent model for all parameters, i.e.,{p t,φ t, f t }. The number of groups and sampling intervals would have been entered as seen in the POPAN formulation. The survival and catchability PIMs mimic those for the POPAN formulation. In the Link-Barker formulation, there are K 1 = 7 population recruitment parameters with a standard PIM:

12.4.2. Link-Barker and Pradel-recruitment formulations 12-26 Because the population recruitment value is not limited to lie between 0 and 1 (for example, the recruitment value could exceed 1), the Parameter Specific Link functions should be specified when models are run: * Any of the link functions can be used for the catchability and survival parameters (although the logit and sin link are most common), but either the log or the identity link function should be used for the population recruitment parameters. There are no restrictions that the recruitment parameters sum to 1 over experiment so the Mlogit link should not be used. Run the model and append it to the browser. As in the POPAN formulation, the fully time-dependent Link-Barker and Pradel-recruitment formulations suffer from confounding. If you examine the β parameter estimates, and the estimated SE (shown at the top of the next page), there are several clues that confounding has taken place: * An undocumented feature of MARK is that it will use the log link for the recruitment parameter if you specify a logit or sin link in the radio buttons.

12.4.2. Link-Barker and Pradel-recruitment formulations 12-27 As in the POPAN formulation, survival in the last interval and catchability at the last sampling occasion are confounded. The corresponding β parameters are very large on the logit scale with standard errors that are either zero or very large. Similarly, p 1 cannot be estimated and itsβstandard error is very large. Finally, because the first and last catchabilities cannot be estimated, neither can population size (despite population size not being explicitly in the model) and so the recruitment parameter (the f i values) based on occasion 1 ( f 1 ) or terminating with occasion K ( f K 1 ) cannot be estimated either. We notice that the standard errors for these recruitment parameters are nonsensical. The user must be very careful to count parameters carefully and to see if MARK has detected the correct number of parameters. There are 8 sampling occasions. The fully time-dependent model has, on

12.4.2. Link-Barker and Pradel-recruitment formulations 12-28 the surface, 8 capture parameters, 7 survival parameters, and 7 recruitment recruitment parameters for a total of 22 parameters. However, as shown in Table 12.3, there are two parameters lost to confounding which gives a total of 20 parameters that can be estimated. The number of parameters reported in the results browser may have to be modified manually. As in the POPAN formulation, models where constraints are placed on the initial and final catchabilities can resolve this confounding. Because roughly the same effort was used in all sampling occasions, the model {p,φ t, f t } seems appropriate. Adjust the PIM for the recapture probabilities to be constant over time, re-run the model (don t forget to use the Parameter Specific Link functions), * and append the results to the browser. Let s look at the parameter estimates: Goodness-of-fit analysis is done using the RELEASE suite as seen in the POPAN fit and as explained in Chapter 5. The number of parameters that can be estimated is now 15 being composed of 1 capture parameter, * CAUTION: When we ran this model with the logit link for theφ s, oneφconverged to a value of 1; we would recommend that the sin link be used.

12.4.2. Link-Barker and Pradel-recruitment formulations 12-29 7 survival parameters, and 7 recruitment parameters. If you compare the estimates of p andφ between the Link-Barker formulation and the POPAN formulation they are identical (except for rounding errors). Note that because of unequal time intervals, estimates ofφ i are on a per-unit basis. The actual survival in the first and last interval must be obtained by raising the reportedφ s to the 1.5 th power (corresponding to the 1.5 week interval). The PENTs from POPAN and the f s from the Link-Barker are not directly comparable. However, the following equivalents are noted: ( ˆ f LB 1 ) 1.5 = 0.252 1.5 = 0.127 = ˆ f LB 2 = 1.644 = ˆ f LB 3 = 0.211 = ˆB POPAN 2 ˆN POPAN 2 POPAN ˆB 3 ˆN POPAN 3 ˆB POPAN 1 ˆN POPAN 1 = 110.39 67.20 = 36.77 171.72 Similarly, estimates of population growth are also equivalent:... ( f ˆLB 1 ) 1.5 LB + ( ˆφ 1 )1.5 = 0.583 1.5 + 0.252 1.5 = 0.571 = ˆ f LB 2 + ˆφ LB 2 = 1.644 + 0.928 = 2.57 =... ˆN POPAN 3 ˆN POPAN 2 = 14.92 117.034 ˆN POPAN 2 ˆN POPAN 1 = 67.20 117.03 = 0.574 = 171.72 67.20 = 2.55 If you examine the results browser (after any changes for the actual number of parameters that are estimated): you will also see that the AIC c values between these two models matches very closely with the difference in the POPAN formulation. The differences in AIC c between the two formulation are artifacts of the different number of parameters estimated and the small sample correction applied to the AIC. If the actual AIC values from the two formulations are compared, the difference in AIC are nearly identical because the two formulations are simply re-parameterizations of the same models for modeling the marked animals and only differ in estimating the super-population size. As Sanathanan (1972, 1977) showed, the conditional approach of Link and Barker (2005) is asymptotically equivalent to the full likelihood approach of Schwarz and Arnason (1996). For example, in the table below, the log-likelihood and AIC (before small sample corrections are applied) were extracted from the model outputs. The differences in the log-likelihoods and the AIC among the models in different formulations

12.4.3. Burnham Jolly-Seber and Pradel-λ formulations 12-30 are nearly the same. POPAN formulation Link-Barker formulation Model 2 log L # parms AIC Model 2 log L # parms AIC {p,φ t, b t } 489.94 16 521.94 {p,φ t, f t } 1176.74 15 1206.74 {p,φ, b t } 507.25 10 527.25 {p,φ, f t } 1193.99 9 1211.99 {p t,φ t, b t } 486.80 21 528.80 {p t,φ t, f t } 1173.69 20 1213.69 The two group case can be fit in a similar fashion as seen in the POPAN example. However, it is not clear exactly which models should be compared because the f i parameters in the Link-Barker formulation depend both upon the NET number of new births, but also upon the population size at occasion i which depends upon the pattern of previous births and survival probabilities. Consequently, even if the same birth pattern occurred between the two groups, differences in survival probabilities could result in differences in patterns of the f s. Fortunately, it appears the survival probabilities are roughly constant between groups, so a comparison of models{p g,φ t, f g t } vs{p g,φ t, f t } is a valid test of the hypothesis of equal return patterns for adults and jacks. Fitting these two models to the two groups is left as an exercise for the reader. 12.4.3. Burnham Jolly-Seber and Pradel-λ formulations The Burnham model The Burnham JS model does not model entrants directly, but rather parameterizes changes in population size using population growth. In the case of the spawning salmon, this would correspond to the increase in the number of spawning salmon at occasion i + 1 relative to occasion i. The Burnham JS model is selected using the Jolly-Seber radio button. The same data file as for POPAN can be used. Again let us start with fitting a model just to the adults over 8 sampling occasions with unequal sampling intervals (see the screen shots in section 12.4.1 for details). Again, start by fitting the fully time-dependent model: As in POPAN (and all JS formulations) there are (K 1) = 7 survival parameters, K = 8 capture

12.4.3. Burnham Jolly-Seber and Pradel-λ formulations 12-31 probabilities. In the Burnham model, there is a single initial population size parameter per group and (K 1) = 7 population growth parameters. We were unable to get the Burnham Jolly-Seber model to converge for any of the models considered in this chapter. The MARK help files state that This model can be difficult to get numerical convergence of the parameter estimates. Although this model has been thoroughly checked, and found to be correct, the program has difficulty obtaining numerical solutions for the parameters because of the penalty constraints required to keep the parameters consistent with each other. Bummer... * The Pradel-λ model The Pradel-λ formulation (considered in much more detail in Chapter 13) can be conveniently obtained by switching data types from any of the JS formulations (or can be entered directly from the initial screen of MARK as discussed in Chapter 12). The number of groups and sampling intervals would have been entered as seen in the POPAN formulation. Let us begin by modeling only the adult salmon. *...dude. C. Schwarz has clearly spent too much time on the wet coast - E. G. Cooch, pers. obs. As noted earlier, this convergence problem disappears for some models if simulated annealing is used for the numerical optimization. - J. Laake & E. G. Cooch, pers. obs.

12.4.3. Burnham Jolly-Seber and Pradel-λ formulations 12-32 Begin by fitting a fully time-dependent model. The PIMs for survival and catchability mimic those seen in earlier sections; the PIM for the population growth parameter has (K 1) = 7 entries: Because the population growth parameter is not limited to lie between 0 and 1 (for example, the growth rate could exceed 1), the Parameter Specific Link functions should be specified when models are run: * Any of the link functions can be used for the catchability and survival parameters (although the logit and sin link are most common), but either the log or the identity link function should be used for the population growth parameters (this is discussed in detail in Chapter 13). There are no restrictions that the growth parameters sum to 1 over experiment so the Mlogit link should not be used. Because there are more than 20 parameters, we need to press the More button to specify the link function for the two remainingλvalues. * A hidden feature of MARK is that it will use the log link for the growth parameter if you specify a logit or sin link in the radio buttons.

12.4.3. Burnham Jolly-Seber and Pradel-λ formulations 12-33 Run the model and append it to the browser. As in the POPAN formulation, the fully time-dependent Pradel-λ formulation suffers from confounding. If you examine the parameter estimates, there are several clues that confounding has taken place. The standard errors for some parameters are enormous or the standard error are zero. Survival in the last interval and catchability at the last sampling occasion are confounded. Similarly,λ 1 is confounded with p 1.

12.4.3. Burnham Jolly-Seber and Pradel-λ formulations 12-34 Consequently, not all of the real parameter estimates are usable: The user must be very careful to count parameters carefully and to see if MARK has detected the correct number of parameters. There are 8 sampling occasions. The fully time-dependent model has, on the surface, 8 capture parameters, 7 survival parameters, and 7 growth parameters for a total of 22 parameters. However, there are two parameters lost to confounding which gives a total of 20 parameters that can be estimated. The number of parameters reported in the results browser may have to be modified manually. As in the POPAN formulation, models where constraints are placed on the initial and final catchabilities can resolve this confounding. Because roughly the same effort was used in all sampling occasions, the model {p,φ t,λ t } seems appropriate. Adjust the PIM for the recapture probabilities to be constant over time, re-run the model (don t forget to use the Parameter Specific link functions or let MARK automatically use the log link for theλ parameters).

12.4.3. Burnham Jolly-Seber and Pradel-λ formulations 12-35 This gives the final estimates: The number of parameters that can be estimated is now 15 being composed of 1 capture parameter, 7 survival parameters, and 7 λ parameters. Note that because of unequal time intervals, estimates ofφ i are on a per-unit basis. The actual survival in the first and last interval must be obtained by raising the reportedφ s to the 1.5 th power (corresponding to the 1.5 week interval). Compare the estimates from the Pradel-λ formulation to those from the POPAN or Link-Barker formulation. Estimates of survival are similar at sampling occasions 1, 2, and 3 differing only by roundoff error. However, the estimate ofφat sampling occasion 4 differs considerably among the models. Indeed, the Pradel-λ formulation is not even consistent as ˆλ 4 = 0.86 < ˆφ 4 = 1.0! The POPAN formulation estimated that there was no recruitment between sampling occasion 4 and 5 (it was constrained so that estimates of recruitment cannot be negative); the Link-Barker model also estimated the recruitment parameter to be 0 (it is also constrained so that it cannot be negative). However, there are no constraints in the Pradel-λ model thatλmust be at least as great as survival. The estimates ofλalso don t appear to be consistent with the results from POPAN when estimates of population size are compared. However, this discrepancy is explained by the different ways in which losses-on-capture are incorporated into the estimates of population size and growth between the two formulations. The model{p,φ,λ t } can also be fit and the results appended to the results browser. The estimates from this model are:

12.5. Example 2 - Muir s (1957) female capsid data 12-36 Again, the estimates are not internally consistent as the population growth rate estimates sometimes fall below the common survival probability. These would indications that there was little or no recruitment in these intervals. The results browser tells the same story as in the other formulations - strong support for the model with constant catchability and time varying survival and population growth. The two group models can also be fit in similar fashion as in the other formulations. However, there is again the question of which models comparisons are a sensible choice. Because the λ values are population growth on a per capita basis and include both survival and recruitment, models with groupand time-dependence in λ may not be indicative of changes in recruitment between the two groups. 12.5. Example 2 - Muir s (1957) female capsid data This second example uses the Muir (1957) data on a population of female black-kneed capsid (Blepharidopterus angulatus) that was originally analyzed by Jolly (1963) * and Seber (1965). According to the British Wildlife Trust website a black-kneed capsid is a green insect about 15 mm long that lives on orchard trees (particular apples and limes). It is a predatory insect that is beneficial to orchard owners as it feeds on red spider mites which cause damage to fruit trees. It apparently makes a squawk by rubbing the tip of its beaks against its thorax and will stab people with its beak if handled. Thirteen successive samples at alternating 3- and 4-day intervals were taken. The population is open as deaths/emigration and births/immigration can occur. The raw history data is located in the file capsid.inp. A portion of the histories appears below: 0000000000001 47; 0000000000010 36; 0000000000011 12; 0000000000100 30; 0000000000101 8;... There is only one group (females) and the individual histories have been grouped. * The data was subsequently reanalyzed in Jolly (1965) http://www.wildlifetrusts.org We suspect that it was impossible to attach individually numbered tags to these small insects and some sort of batch marking scheme (e.g., color dots) was used. This batch marking scheme would enable the history of each insect to be followed, but individuals with the same history cannot be separately identified

12.5. Example 2 - Muir s (1957) female capsid data 12-37 The summary statistics are: Occasion t i n i m i u i R i r i z i 1 0 54 0 54 54 24 0 2 3 144 10 134 143 83 14 3 7 166 39 127 166 71 58 4 10 203 56 147 202 71 73 5 14 186 54 132 185 76 90 6 17 197 66 131 196 92 100 7 21 231 97 134 230 102 95 8 24 164 75 89 164 95 122 9 28 161 101 60 160 69 116 10 31 122 80 42 122 55 105 11 35 118 74 44 117 44 86 12 38 118 70 48 118 35 60 13 42 142 95 47 142 0 0 There are a few losses on capture at some of the sampling occasions where n i R i. The data are input into MARK in the usual fashion. The time intervals are set to three and four day intervals in the usual fashion:

12.5.1. POPAN formulation 12-38 12.5.1. POPAN formulation We begin by fitting the fully time-dependent model {p t,φ t, b t } in the usual fashion using PIMs: The model is run, again selecting the parameter specific link functions:

12.5.1. POPAN formulation 12-39 and specifying the Mlogit(1) link-function for the PENT parameters, and the lo g link-function for the super-population size (N). The resulting model output is: There are a total of 36 parameters (13 capture probabilities; 12 survival probabilities; 13 entry probabilities * 1 super-population size; less 2 confounded parameters at the start and end of the experiment and * Don t forget that MARK will show only the later 12 PENTs in the PIM and output because it never allows you to do anything with the b 0 parameter.

12.5.1. POPAN formulation 12-40 less 1 restriction that the PENTs must sum to 1). The results browser only shows 30 parameters because some of the estimated PENTs, p s andφ s are estimated to be 0 or 1. * The number of parameters in the results browser should be reset to 36 in the usual fashion. Because of the differing time intervals, the estimated survival rates (φ s) are the survival probabilities per day. They also differ from the estimates found in Jolly (1965) because they are also constrained to lie between 0 and 1. For example, Jolly (1965) estimated that the survival probability between the second and third sampling occasion was 1.015. The estimated population sizes (shown at the top of the next page) are found in the derived parameters (accessed using Output Specific Model Output Parameter Estimates Derived Estimates ): * This may be clearer if the β estimates table is also examined.

12.5.1. POPAN formulation 12-41 Because of confounding and non-identifiability at the start and the end of the experiment in the fully time-dependent model, some estimates cannot be used. Model with constant p and/or constantφ per day may also be tenable and are fit in the usual way using the PIMs. Models with constant b s don t really have any sensible biological interpretation and should not be fit. A model for the emergence curve may be more sensible and this could result in predictions about the number of new entrants over time that has an early peak and tends to tail off over time. The number of parameters estimated by MARK should be checked and adjusted. The final results table is: It appears that virtually all support lies on the model{p,φ t, b t }. Because the capture probabilities are constant over time, there is no longer any problem with confounding or non-identifiability at the start or end of the experiment.

12.5.1. POPAN formulation 12-42 The final estimates of the basic parameters and the derived parameters are: The estimated super-population size is interpreted as the total number of capsids ever present in the experiment and does not represent the number present at any particular point in time. The population seems to peak at about sampling occasion 4, and then gradually tapers off because of deaths/emigration and fewer new insects entering the population of interest.

12.5.2. Link-Barker and Pradel-recruitment formulations 12-43 12.5.2. Link-Barker and Pradel-recruitment formulations Both the Link-Barker and Pradel-recruitment parameterize new entrants to the population using the f i parameter representing the numbers of new recruits in the interval per member of the population alive at time i. Because this dataset includes losses-on-capture, the Pradel-recruitment model cannot be used, and the Link-Barker formulation will be used. The data are entered as shown above don t forget to set the alternating 3- and 4-day time intervals. The fully time-dependent model {p t,φ t, f t } is fit using PIMs in the usual fashion: This model is run in the usual fashion. The recruitment parameter should have a lo g link-function specified as this parameter can exceed the value of 1.

12.5.2. Link-Barker and Pradel-recruitment formulations 12-44 The results table is: There are 35 parameters (13 capture probabilities; 12 survival probabilities; 12 recruitment probabilities; less 1 non-identifiable parameter at the end of the sampling chain whereφ 12 p 13 can only be estimated; and less 1 non-identifiable parameter at the start of the sampling chain.) If the number of parameters in the results table differs, it should be changed in the usual fashion. The estimates from this model are shown at the top of the next page. The estimated survival probabilities and recruitment parameters are on a per day basis. The estimates of p andφare comparable to the POPAN estimates except for some minor differences at the start of the sampling chain. These are artifacts of the different confounding at the start of the sampling chain between the two formulations. The

12.5.2. Link-Barker and Pradel-recruitment formulations 12-45 estimated recruitment parameters indicate the number of new recruits per member of the population. Estimates of the actual population size are not available. Simpler models for p and φ can be fit in the usual fashion. It may be actually biologically sensible to fit a model with constant f over time as this is the recruitment per existing member. Even if the population is declining over time, perhaps the recruitment is constant over time. The results table for these models is: Virtually all support again lies with the model with constant catchability{p,φ t, f t }. Because capture probabilities were constant over time, all parameters are now estimable. The final estimates are:

12.5.3. Burnham Jolly-Seber and Pradel-λ formulations 12-46 The estimates of p andφare comparable to those from the POPANformulation except at the start of the sampling chain. This may be an artifact of convergence to a local minimum by MARK. The estimates of recruitment can be matched to that from POPAN. For example, Now ˆ f LB 12 = 0.54765, ˆB POPAN 12 = 43.83, and ˆN POPAN 12 = 486.02. ˆB POPAN 12 ˆN POPAN 12 = 43.83 486.02 = 0.090 = ( f ˆLB 12 )4 = 0.548 4 = 0.090 12.5.3. Burnham Jolly-Seber and Pradel-λ formulations The Burnham-Jolly-Seber and the Pradel-λ formulations parameterize new recruits to the population indirectly by estimating population growth (λ) representing the population size at time i + 1 relative to the population size at time i. The growth process is the net effect of both survival and recruitment. The data are entered in the usual fashion - don t forget to set the alternating 3- and 4-day intervals. The Burnham-Jolly-Seber formulation again has difficulty in convergence for this example and so is not run against this data.

12.5.3. Burnham Jolly-Seber and Pradel-λ formulations 12-47 The fully time-dependent Pradel-λ model {p t,φ t,λ t } is fit using PIMs in the usual fashion: This model is run in the usual fashion. The population growth parameter should have a lo g link-function specified as this parameter can exceed the value of 1.

12.5.3. Burnham Jolly-Seber and Pradel-λ formulations 12-48 The results table is: There are 35 parameters (13 capture probabilities; 12 survival probabilities; 12 growth rates; less 1 nonidentifiable parameter at the end of the sampling chain whereφ 12 p 13 can only be estimated; and less 1 non-identifiable parameter at the start of the sampling chain.) If the number of parameters in the results table differs, it should be changed in the usual fashion. The estimates from this model are:

12.5.3. Burnham Jolly-Seber and Pradel-λ formulations 12-49 The estimated survival probabilities and growth rates are on a per day basis. The estimates of p andφ are comparable to the POPAN estimates except for some minor differences at the start of the sampling chain. These are artifacts of the different confounding at the start of the sampling chain between the two formulations. The estimated growth parameters indicate ratio of the estimated population size at successive sampling intervals on a per unit time basis. Estimates of the actual population size are not available directly, but as illustrated in previous examples can be derived if needed. Simpler models for p andφcan be fit in the usual fashion. It may be actually biologically sensible to fit a model with constantλover time if the population is roughly constant over time. Because the growth rate includes both survival and recruitment, models where growth is constant over time, but survival is not, are not usually fit as it is difficult to believe that changes in recruitment will exactly balance changes in survival to keep the population at a constant level. The results table for these simpler models is: Virtually all support again lies with the model with constant catchability {p,φ t,λ t }. Because capture rates were constant over time, all parameters are now estimable. The final estimates are: The estimates of p andφare comparable to those from the POPAN formulation. There are only a few sampling occasions where the estimates of population growth are inconsistent with estimates of survival, but the differences are minor. The estimates of population can be matched to that from POPAN. For example, ˆN POPAN 12 = 486.02, and ˆN POPAN 13 = 529.86. ˆλ PL 12 = 1.0219978,